The exponential growth of images makes the traditional single machine image retrieval face the problem of slow retrieval speed, poor concurrency and low image accuracy when dealing with large-scale images. According to that the image feature files are small files, this paper proposed to properly merge the small files, and then put store them on the distributed file system HDFS of Hadoop. It achieved rapidly store and read massive image data. In order to adapt to the large-scale image retrieval, this paper proposed to binarize Fisher vector of images and use MapReduce parallel programming model to realize parallel image retrieval based on binary Fisher vector and SIFT. Experiments on INRIA Holidays dataset, Kentucky dataset and Flicker1M dataset show that the method is scalable, can achieve better retrieval accuracy, effectively reduce the retrieval time and improve the retrieval speed. It is a highly efficient large-scale image storage and retrieval method.%图像数据的指数型增长使得传统单机的图像检索在处理大规模图像时面临着检索速度慢、并发性差、检索准确率低的问题.由于图像特征文件都是小文件,本文提出将图像特征小文件进行适当的合并后存储于Hadoop的分布式文件系统HDFS中,实现大规模图像的快速存储和读取;为了适应大规模的图像检索,对图像Fisher向量进行二值化处理,并利用MapReduce并行编程模型实现基于二值Fisher向量和SIFT(Scale Invariant Feature Transform)特征的并行检索.在INRIA Holidays数据集、Kentucky数据集和Flicker1M数据集上的实验结果表明该方法扩展性强,能够取得较好的检索准确率,有效减少检索时间,提高检索速度,是一种高效的大规模图像存储和检索的方法.
展开▼
